Deep research and synthesis

Several passes over web and academic sources, each claim cross-checked against a second source, ending in a written report with its citations, saved to a file and summarised into memory.

This is the expensive path. For a question two or three searches would settle, use web-research instead: it is the same discipline at a tenth of the cost, and choosing it correctly is part of doing this well.

0. Declare the mission before spending anything

research_mode with mode set, and a short reason. Call it BEFORE the first search, as soon as the user asks for something thorough, deep or exhaustive, or as soon as you realise a single lookup will not be enough. It activates the long-running research behaviour: budgets, pacing, and the right expectations about how many passes are allowed.

Declaring it late is the usual mistake. Three shallow searches happen first, the answer is thin, and the mode is switched on when the work is nearly over.

Then check what you already know, which is free:

• memory_search with query set to the subject. This may have been researched before.
• session_search with query runs a full-text search across PAST sessions and returns matching excerpts. Use it when the user says "we looked at this already", or when the subject feels familiar. It is far cheaper than repeating the research.

1. Lock the question

Before any search, write down, for yourself:

• the research question, in ONE sentence;
• three to five domain keywords, in the vocabulary the sources use, not yours;
• the target: how many sources, how recent, and where the output goes.

Defaults when the user did not say: five to fifteen sources, an eighteen month recency window, a report written to a file. A question that has not been reduced to one sentence produces a report about several subjects and answers none of them.

2. Search, broad then narrow

web_search(query="<keywords> survey OR review 2026", num_results=10)
web_search(query="<keywords>", allowed_domains=["arxiv.org"], num_results=10)
web_deep_search(query="<one specific sub-question>")

num_results defaults to 8 and is capped at 15. allowed_domains and blocked_domains are arrays: use the first to pin a pass to a primary source, the second to cut a content farm that keeps winning the ranking.

Run the passes in PARALLEL, one angle per call. They are read-only, so several fit in one message. Collect arXiv ids, titles, official docs. De-duplicate as you go, not at the end.

web_deep_search returns snippets AND the full text of the top results. It is slower and worth it on a sub-question that shallow search keeps answering with noise. It is not the tool for the first, wide pass.

3. Read the sources, not the snippets

web_fetch(url="https://arxiv.org/abs/<id>")
web_fetch(url="https://arxiv.org/pdf/<id>")

web_fetch takes ONE url, as a string, and extracts a PDF served over HTTP by itself. A long page arrives in chunks: the output states the total size and the offset for the next one. If the section you need is not in the first chunk, read on with offset.

When a server refuses the fetch, download and read from disk instead:

curl -sL "https://arxiv.org/pdf/<id>" -o "C:\\tmp\\paper.pdf"

then read_extract with path set to that absolute path. read_extract reads LOCAL PDF, .txt and .md files only; it has no urls argument and no field selection.

Discard as you read: no date, no author, a preprint over two years old with no citations, a page that restates the question without answering it.

Record every decisive fact with finding the moment you read it, passing fact and source. The findings ledger survives context compaction, and a multi-pass mission WILL be compacted. A fact you did not record can be gone before you write the report.

4. Cross-check adversarially

For each claim that carries weight:

• find a second, INDEPENDENT source, or mark the claim [UNVERIFIED] in the report;
• two pages repeating the same press release are one source, not two;
• when sources disagree, write both positions and say the question is open. Picking a side silently is the failure this step exists to prevent.

5. Write the report

# <Topic>, research report, <YYYY-MM-DD>

## Answer
<the answer to the one-sentence question, first, in two or three sentences>

## Key findings
- <finding> [<title>, <url>]

## Disagreements and gaps
- <claim A> against <claim B>, unresolved because <reason>
- <what could not be established, and what was tried>

## Sources
| # | Title | URL | Date | Kind |

## Open questions

Then save it, and leave a trace that survives the session:

file_write(path="C:/dev/research/<topic>_report.md", content="<the report>")
memory_write(node_id="research.<topic>", content="<two sentences and the file path>")

memory_write takes node_id and content. It has no key or value. node_id is dotted and snake_case (research.world_models), never a slash path. Use an absolute path in file_write: a relative one lands in the server's working directory, not the user's.

Traps

• Answering from abstracts. An abstract states what the authors hoped. For a claim the report leans on, read the results and the limitations sections.
• arXiv versioning. arxiv.org/abs/1706.03762 is the LATEST version and can differ substantially from what you read last month. Cite the version you actually read, with its v suffix.
• Recency mistaken for relevance. The most recent paper is not the most important one. Check citation counts before ranking.
• The single-source finding. One page saying something is a claim. Two pages copying each other is still one claim.
• Reporting the research instead of the answer. The user asked a question. Lead with the answer; the method belongs below it.

Failure modes

Every result is the same syndicated article. You found a press release. Search for what it is based on: the paper, the filing, the repository, the official announcement.

A source is paywalled or returns 403. An obstacle, not a conclusion. Try web_fetch with render: true, then the Wayback Machine, then another source. Only report the angle as blocked after all of those.

Context ran out mid-mission and the findings are gone. They were never written with finding. Restart the pass, recording as you read this time.

The report has ten sources and no answer. The question was never reduced to one sentence. Go back to step 1 and write it, then re-read what you already collected against it: most of it usually answers something else.